面向复杂动作的运动状态转移识别模型

Motion State Transition Recognition Model for Complex Actions

基于惯性传感器的人类活动识别一直以来都是研究热点,大多数的研究都是基于单个动作的分类识别,面对日常生活中具有多个运动状态且连续的动作信号时,很难对其中存在的运动状态进行精确分割和识别.论文提出一种面向复杂运动的运动状态转移识别模型,将人类日常动作分为动态动作、静态动作和过渡动作,并对这几类动作进行分割和识别.通过滑动窗口分割法进行预分割,分割周期性强的动态动作和易识别的静态动作,获取动态动作分割结果和静态动作先验.根据静态动作和过渡动作变化率的差异,设置阈值来区分静态动作和过渡动作的边界.最后,根据静态动作状态和过渡动作状态持续时间不同的逻辑关系建立有限状态机,利用静态动作先验将过渡动作识别出来.该模型在UCI公开的数据集上随机抽取了五组测试样本,整体分割识别准确率最高达到98.25%,最低也达到95%以上.

Human activity recognition based on inertial sensors has always been a research hotspot.Most of the research is based on the classification and recognition of a single action.When faced with multiple motion states and continuous action signals in daily life,it is difficult to identify the existing ones.The motion state is accurately segmented and recognized.The paper proposes a recognition model for complex motion-oriented motion state transition,which divides human daily actions into dynamic actions,static actions and transitional actions,and divides and recognizes these types of actions.Pre-segmentation is performed by sliding window segmentation method to segment dynamic actions with strong periodicity and static actions that are easy to recognize,and obtain dynamic action segmentation results and static action priors.According to the difference in the rate of change between static actions and transition actions,a threshold is set to distinguish the boundary between static actions and transition actions.Finally,a finite state machine is established according to the logical relationship between the static action state and the transition action state duration,and the transition action is identified by the static action a priori.The model randomly selected five test samples on the data set disclosed by UCI,and the overall segmentation recognition accuracy reached 98.25%at the highest and 95%at the lowest.